Claude Code ships a built-in agent-creation wizard at /agents. It walks you through one agent at a time: type, description, prompt, model, color, tools, memory, location, confirm. The system prompt it uses to generate the agent for you is AGENT_CREATION_SYSTEM_PROMPT in the binary, and it has a lot of opinions about what a good subagent looks like (identifier rules, Use this agent when ... descriptions with <example> blocks, second-person prompts, "recently written code" defaults for reviewers, etc.).

That wizard is fine for one-off agents. But if you sit down in a fresh repo and want a useful set of subagents, you don't want to run the wizard six times and answer the same context questions. You want something that reads the repo, figures out what a good set of agents would be, and writes them all in one shot.

So I built /agentify. It scans the project (manifests, CI, CLAUDE.md, source layout, existing .claude/ setup), proposes a set of agents tailored to what it finds, and writes them to .claude/agents/*.md after you confirm. The authoring discipline mirrors AGENT_CREATION_SYSTEM_PROMPT and the validation mirrors validateAgent.ts so the output is the same shape Claude Code's own wizard produces.

Given a project directory (defaults to your current working directory):

1. Resolve the project root via pwd / git rev-parse --show-toplevel.
2. Scan the codebase: CLAUDE.md, AGENTS.md, README*, package manifests, CI configs, source layout, existing .claude/agents/ and .claude/skills/.
3. Propose a set of subagents. Each one points to evidence from the scan (a real workflow, file, command, or convention). Minimum one. Zero is a valid outcome if the project doesn't justify any. No padding to hit a count.
4. Wait for your confirmation. No files written before you say yes.
5. Write each agent to <project>/.claude/agents/<name>.md with canonical frontmatter (name, description, tools, model, effort, color, memory in that order) and a second-person system prompt that references real paths and conventions from the scan.
6. Validate against the real rules: name regex ^[a-zA-Z0-9][a-zA-Z0-9-]*[a-zA-Z0-9]$ (length 3-50), enum-checked model/color/memory/permissionMode/isolation, description starting with Use this agent when ... with at least one <example> block.

Add the marketplace and install the plugin in Claude Code:

/plugin marketplace add 0xMH/claude-agentify
/plugin install agentify@0xMH/claude-agentify

Or test locally during development:

claude --plugin-dir ./path/to/claude-agentify

From inside a project:

/agentify

Or pass a path or a brief:

/agentify ~/code/my-app
/agentify make me a migration checker and a release auditor

After it writes files to .claude/agents/, you can use the agents three ways:

1. Auto-delegation: the main Claude reads each agent's description and routes matching tasks to it. The <example> blocks in the description are what drives this, so they matter.
2. Explicit: Use the <agent-name> subagent to ...
3. /agents: list, view, edit, invoke from the built-in UI.

Newly written agents may need a session restart to load.

The built-in /agents wizard creates one agent per run and asks you for every dimension. agentify:

• Reads the project and proposes a tailored set, so you confirm once instead of running the wizard N times.
• Forces evidence per agent. If a proposed agent can't point to something in the scan, it gets dropped. No generic code-reviewer filler.
• Restricts tools by role automatically (read-only roles do not get Write/Edit).
• Mirrors the AGENT_CREATION_SYSTEM_PROMPT discipline (identifier rules, Use this agent when ... descriptions with <example> blocks, second-person prompts, code-review "recently written" default, memory section pattern when memory: is set).
• Writes the canonical field set (name, description, tools, model, effort, color, memory) in the same order Claude Code's own writer uses.

Agents written by agentify follow this structure:

---
name: project-role
description: "Use this agent when ... <example>Context: ...\nuser: \"...\"\nassistant: \"...\"\n<commentary>...</commentary></example>"
tools: Read, Glob, Grep
model: inherit
color: blue
---

You are ...

Field rules (enforced by Claude Code's own validator):

• name (required): regex ^[a-zA-Z0-9][a-zA-Z0-9-]*[a-zA-Z0-9]$, length 3-50, must equal filename without .md.
• description (required): YAML double-quoted, newlines escaped as \n. Should start with Use this agent when ... and embed <example> blocks.
• tools (optional): comma-separated. Omit the field entirely to inherit all tools.
• model (optional): sonnet | opus | haiku | inherit.
• color (optional): red | blue | green | yellow | purple | orange | pink | cyan.
• memory (optional): user | project | local.
• effort (optional): enum string or positive integer.

Reader-only fields parsed if present: disallowedTools, permissionMode, maxTurns, isolation, background, initialPrompt, skills, mcpServers, hooks.

MIT